You are on page 1of 4

Unversitat Pompeu Fabra Kurt Schmidheiny Clustering in the Linear Model 2

Short Guides to Microeconometrics October 2008

observations in cluster g, N = g Ng is the total number of observations,
yig is the dependent variable, xig is a (K +1)-dimensional row vector of K
explanatory variables plus a constant, β is a (K + 1)-dimensional column
Clustering in the Linear Model vector of parameters, and εig is the error term.
Stacking observations within a cluster, we can write

yg = Xg β + εg
1 Introduction where yg is a Ng × 1 vector, Xg is a Ng × (K + 1) matrix and εg is is a
Ng × 1 vector. Stacking observations cluster by cluster, we can write
This handout extends the handout on ”The Multiple Linear Regression
model” and refers to its definitions and assumptions in section 2. It y = Xβ + ε
relaxes the homoscedasticity assumption (A5a) and allows the error terms
where y = [y1 ... yG
] is N × 1, Xg is N × (K + 1) and εg is N × 1.
to be heteroscedastic and correlated within groups or so-called clusters.
The data generation process (dgp) is fully described by the following
It shows in what situations the parameters of the linear model can be
set of assumptions:
consistently estimated by OLS and how the standard errors need to be
corrected. A1: Linearity
The canonical example (Moulton 1990) for clustering is a regression
yi = xig β + εig and E(εig ) = 0
of individual outcomes (e.g. wages) on explanatory variables of which
some are observed on a more aggregate level (e.g. employment growth A2: Independence
on the state level).
c) (Xg , yg )G
g=1 independently distributed
Clustering also arises when the sampling mechanism first draws a ran-
dom sample of groups (e.g. schools, households, towns) and than surveys A2c means that the observations in one cluster are independent from the
all (or a random sample of) observations within that group. Stratified observations in all other clusters.
sampling, where some observations are intentionally under- or oversam- A3: Strict Exogeneity
pled asks for more sophisticated techniques.
a) εig |Xg ∼ N (0, σig )

2 The Econometric Model b) εig ⊥ Xg and E(εig ) = 0 (independent)

c) E(εig |Xg ) = 0 (mean independent)
Consider the multiple linear regression model
d) Cov(Xg , εig ) = 0 and E(εig ) = 0 (uncorrelated)
yig = xig β + εig
Note that the error term εig is assumed unrelated to the explanatory
where observations belong to a cluster g = 1, ..., G and observations variables (Xg ) of all observations within its cluster.
are indexed by i = 1, ..., Ng within their cluster. Ng is the number of

Version: 31-10-2008, 18:37

3 Short Guides to Microeconometrics Clustering in the Linear Model 4

A4: Identifiability 3 A Special Case: Cluster Specific Random Effects

rank(X) = K + 1 < N
Suppose as Moulton(1986) that the error term εig consists of a cluster
A5: Error Variance specific random effect αg and an individual effect νig
c) V (εig |Xg ) = σig < ∞, for all i, g εig = αg + νig
Cov(εig , εjg |Xg ) = ρijg σig σjg < ∞, for all i = j, g
Assume that the individual error term is homoscedastic and independent
A5c means that the error terms are correlated within clusters (clustered)
across all observations
and have different variances (heteroscedastic).
V (νig |Xg ) = σν2
A6: Variance of explanatory variables Cov(νig , νjg |Xg ) = 0, i = j
a) V (X) = E(X  X) is positive definite and finite
and that the cluster specific effect is homoscedastic and uncorrelated with
b) plim( N1 X  X) = QXX is positive definite and finite the individual effect
V (αg |Xg ) = σα2
The variance-covariance of the vector of error terms in the whole sample
is under A2 and A5 Cov(αg , νig |Xg ) = 0
The cluster specific effect αg is under A3 at least uncorrelated with Xg
Ω = V (ε|X) = E(εε |X)
and can therefore be treated as a random effect:
⎛ ⎞
Ω1 0 · · · 0
⎜ ⎟ Cov(αg , Xg ) = 0 .
⎜ 0 Ω2 · · · 0 ⎟
=⎜⎜ .. .. . . .. ⎟

⎝ . . . . ⎠ The resulting variance-covariance structure within each cluster g is
0 0 · · · ΩG then ⎛ ⎞
σ 2 σ 2 · · · ρσ 2
where, for example, ⎜ ⎟
⎜ ρσ 2 σ 2 · · · ρσ 2 ⎟
Ωg = V (εg |Xg ) = ⎜
⎜ .. .. ... .. ⎟

Ω1 = V (ε1 |X1 ) = E(ε1 ε1 |X1 ) ⎝ . . . ⎠
⎛ ⎞ ρσ 2 ρσ 2 · · · σ2
σ12 ρ12 σ1 σ2 · · · ρ1N1 σ1 σN1
⎜ ⎟ where σ 2 = σα2 + σν2 and ρ = σα2 /(σα2 + σν2 ). In a less restrictive version,
⎜ ρ12 σ1 σ2 σ22 · · · ρ2N1 σ2 σN1 ⎟
=⎜⎜ .. .. ... .. ⎟
⎟ σg2 and ρg are allowed to be cluster specific.
⎝ . . . ⎠
2 Note: this structure is identical to a panel data random effects model
ρ1N1 σ1 σN1 ρ2N1 σ2 σN1 · · · σN 1
with many individuals g observed over few time periods i.
is the variance covariance of the error terms within cluster g = 1.
5 Short Guides to Microeconometrics Clustering in the Linear Model 6

4 Estimation with OLS where V = G−1 Q−1 ΣQ−1 can be consistently estimated as
The parameter β can be estimated with OLS as −1
V̂ = (X  X) Xg eg eg Xg (X  X)
−1 g=1
β̂OLS = (X  X) X y
with eg = yg − Xg β̂OLS .
The OLS estimator of β remains unbiased (under A1, A2c, A3c, A4, This so-called cluster-robust covariance matrix estimator is a gen-
A5c and A6) and normally distributed (additionally assuming A3a) in eralization of Huber(1967) and White(1980).1 It does not impose any
small samples. It is consistent and approximately normally distributed restrictions on the form of both heteroscedasticity and correlation within
(under A1, A2c, A3d, A4, A5c and A6b) in samples with a large number clusters (though we assumed independence of the error terms across clus-
of clusters. However, the OLS estimator is not efficient any more. More ters). We can perform the usual z- and Wald-test for large samples using
importantly, the usual standard errors of the OLS estimator and tests the cluster-robust covariance estimator.
(t-, F -, z-, Wald-) based on them are not valid any more. Note: the cluster-robust covariance matrix is consistent when the
number of clusters G → ∞ and the number of observations per cluster
5 Estimating the Covariance of the OLS Estimator Ng is fixed. In practice this requires a sample with many clusters (50 or
more) and relatively small number of observations per cluster.
The small sample covariance matrix of β̂OLS is under A3c and A5c
Bootstrapping is an alternative method to estimate a cluster-robust
−1  2
−1 covariance matrix under the same assumptions. See the handout on
V = V (β̂OLS |X) = (X  X) X σ ΩX (X  X)
”The Bootstrap”. Clustering is addressed in the bootstrap by randomly
and differs from usual OLS where V = σ 2 (X  X)−1 . Consequently, the drawing clusters g (rather than individual observations ig) and taking
usual estimator V̂ = σ̂ 2 (X  X)−1 is incorrect. Usual small sample test all Ng observations for each drawn cluster. This so-called block bootstrap
procedures, such as the F - or t-Test, based on the usual estimator are preserves all within cluster correlation.
therefore not valid.
With the number of clusters G → ∞, the OLS estimator is asymp- 6 Estimation with Cluster Specific Random Effects
totically normally distributed under A1, A2, A3d, A4, A5c and A6b
√ In the cluster specific random effects model, the error covariance matrix
G(β̂ − β) −→ N 0, Q−1 ΣQ−1
Ω only depends on the two parameters ρ and σ. These two parameters
can be consistently estimated in samples with many clusters. We could
The OLS estimator is therefore approximately normally distributed in
plug these estimates into Ω to estimate the correct covariance V̂ for the
samples with a large number of clusters
OLS estimator β̂OLS .
β̂ ∼ N (β, V ) . 1
Note: the cluster-robust estimator is not clearly attributed to a specific author.
See e.g.
7 Short Guides to Microeconometrics Clustering in the Linear Model 8

However, if we are willing to assume cluster specific random effects, References

we can directly estimate β efficiently using feasible GLS (see the handout
on ”Heteroscedasticity in the Linear Model” and the handout on ”Panel Cameron, A. C. and P. K. Trivedi (2005), Microeconometrics: Methods
Data”). In practice, we can rarely rule out additional serial correlation and Applications, Cambridge University Press. Sections 24.5.
beyond the one induced by the random effect. It is therefore advisable Wooldridge, J. M. (2002), Econometric Analysis of Cross Section and
to always use cluster-robust standard errors in combination with FGLS Panel Data. MIT Press. Sections 7.8 and 11.5.
estimation of the random effects model.
Huber, P. J. (1967), The behavior of maximum likelihood estimates un-
der nonstandard conditions. In: Proceedings of the Fifth Berkeley
7 Implementation in Stata 10.0
Symposium on Mathematical Statistics and Probability. Berkeley, CA:
Stata reports the cluster-robust covariance estimator with the vce(cluster) University of California Press, 1, 221223.
option, e.g.2
Moulton, B. R. (1986) Random Group Effects and the Precision of Re-
webuse auto7.dta gression Estimates, Journal of Econometrics, 32(3): 385-397.
regress price weight, vce(cluster manufacturer)
matrix list e(V)
Moulton, B. R. (1990) An Illustration of a Pitfall in Estimating the
Note: Stata multiplies V̂ with (N − 1)/(N − K) · G/(G − 1) to correct Effects of Aggregate Variables on Micro Units, The Review of Eco-
for degrees of freedom in small samples. nomics and Statistics, 72, 334-338.
We can also estimate a heteroscedasticity robust covariance using a
White, H. (1980), A Heteroscedasticity-Consistent Covariance Matrix
nonparametric block bootstrap. For example,
Estimator and a Direct Test for Heteroscedasticity. Econometrica 48,
regress price weight, vce(bootstrap, rep(100) cluster(manufacturer))
bootstrap, rep(100) cluster(manufacturer): regress price weight

The cluster specific random effects model is efficiently estimated by

FGLS. For example,
xtset manufacturer_grp
xtreg price weight, re

In addition, cluster-robust standard errors are reported with

xtreg price weight, re vce(cluster manufacturer)
There are only 23 clusters in this example dataset used by the Stata manual.
This is not enough to justify using large sample approximations.